SE-LightCNN: A Lightweight Attention-Enhanced CNN for Bell Pepper Disease Classification

Abstract

Global food security depends on agricultural disease control, and manual inspection-based techniques have been used, but they are labor-intensive and prone to human error. In this work, we proposed SE-lightCNN, a lightweight convolutional neural network that can accurately identify phytopathological diseases in real time. Using Squeeze-and-Excitation (SE) blocks for channel-wise feature recalibration on a dataset of 2,375 images, the suggested architecture demonstrates excellent performance with accuracy and precision values of 99.69%, respectively. The model outperforms previously suggested architectures such as ResNet50 and EfficientNet despite having a remarkably small computational footprint of just 159K parameters. In order to apply the theoretical model in the real world, it was deployed as a Flutter based mobile application that provided farmers with offline diagnosis and treatment protocols. This work presents an affordable, open-source scanning solution covering precision agriculture applications on limited edge hardware.